search

VinAIResearch / PhoGPT

PhoGPT: Generative Pre-training for Vietnamese (2023)
Apache License 2.0
744 stars 65 forks source link

PhoGPT-7B5-Instruct has only one word repeated over and over again. #7

Closed hacbachvotu closed 11 months ago

hacbachvotu commented 11 months ago

z4876691733259_7426ad5b582f9fc096dcd0181a9a32fa

PhoGPT-7B5-Instruct has only one word repeated over and over again when i generate! Why does this error appear, everyone? I only have this problem with this model.

datquocnguyen commented 11 months ago

Closed! Following the same question from you in this thread: https://github.com/VinAIResearch/PhoGPT/issues/4 You may want to install pytorch, transformers, and/or run the example code (on Google Colab) with GPU.

hacbachvotu commented 11 months ago

Closed! Following the same question from you in this thread: #4 You might want to install torch, transformers, and/or run the example code (on Google Colab) with GPU.

I have installed all with the latest version and all other models run smoothly except this one.

datquocnguyen commented 11 months ago

https://github.com/VinAIResearch/PhoGPT/issues/4#issuecomment-1807924908 CUDA_VISIBLE_DEVICES=0 python testscript.py would generate "Xin chào. Tôi có thể giúp gì cho bạn?".

hacbachvotu commented 11 months ago

#4 (comment) CUDA_VISIBLE_DEVICES=0 python testscript.py would generate "Xin chào. Tôi có thể giúp gì cho bạn?".

sorry i don't know what did you say! i use oobabooga with GPU 12GB 3080ti.

hacbachvotu commented 11 months ago

z4879313057704_24d48ae641cad8d14ece01ec28238d9c z4879313559849_2d771ddddec3f1ac5cece45da7d08089 everything ok when i use another model. I don't understand why only this model of yours is having problems.

datquocnguyen commented 11 months ago

You may want to enable trust-remote-code.

hacbachvotu commented 11 months ago

You may want to enable trust-remote-code.

z4879988718785_17f448df3586dc037c171268ea8d4d3d z4879989253645_8f641a4783fd3a083f1eb79e9bc60c3a Thanks for you help! It's work with trust-remote-code Flag.

datquocnguyen commented 11 months ago

For the model to work properly, you might need to insert/use the PROMPT template "### Câu hỏi:\n{instruction}\n\n### Trả lời:".

hacbachvotu commented 11 months ago

For the model to work properly, you might need to insert/use the PROMPT template "### Câu hỏi:\n{instruction}\n\n### Trả lời:".

z4880987097593_1f879b805c32207997b9dcd094ee57d2 Let me ask one more question: is it correct to put this prompt in oobooga?

AlexBlack2202 commented 10 months ago

Câu hỏi:\n{instruction}\n\n### Trả lời:"

For the model to work properly, you might need to insert/use the PROMPT template "### Câu hỏi:\n{instruction}\n\n### Trả lời:".

z4880987097593_1f879b805c32207997b9dcd094ee57d2 Let me ask one more question: is it correct to put this prompt in oobooga?

do you know how to correct this?, i have same question